Comparative Local Quality Assessment of 3D Medical Image Segmentations with Focus on Statistical Shape Model-Based...

Area of Science:

• Medical Imaging
• Computer Vision
• Biomedical Engineering

Background:

• Assessing 3D medical segmentation quality across datasets is crucial for identifying systematic errors.
• Current methods lack robust support for comparative evaluation of segmentation quality across multiple 3D images.
• Experts need tools to analyze and compare segmentation performance on datasets containing numerous organ instances.

Purpose of the Study:

• To present a novel system for assessing and comparing 3D medical segmentation quality within datasets.
• To enable the detection and visualization of regions with systematic segmentation quality characteristics.
• To facilitate the identification of both systematic segmentation problems and outlier instances.

Main Methods:

• Developed a novel system for analyzing and visualizing 3D medical segmentation quality across multiple instances.
• Extended a hierarchical clustering algorithm with a connectivity criterion to identify regions with characteristic segmentation quality.
• Combined quality values across datasets to determine regions with consistent segmentation quality across instances.

Main Results:

• The system effectively detects and visualizes regions exhibiting systematic segmentation quality characteristics.
• Experts successfully identified organ regions with systematic segmentation patterns using the proposed system.
• The approach enabled the identification of outlier 3D segmentations with unusual quality characteristics.

Conclusions:

• The novel system significantly improves the assessment and comparison of 3D medical segmentation quality across datasets.
• The enhanced clustering algorithm aids in pinpointing systematic segmentation issues and outliers.
• The approach is applicable to various 3D medical segmentation algorithms, especially those using statistical shape models or landmark correspondences.